基于PCA-PSO-RF的地层结构随钻智能识别方法

Intelligent identification method of formation structure while drilling based on PCA-PSO-RF

摘要:

针对传统钻探取心效率低、信息滞后，且现有随钻识别模型常忽略参数多重共线性与超参数优化而导致分类精度受限的问题，提出一种融合主成分分析（PCA）、粒子群优化（PSO）与随机森林（RF）的地层结构随钻智能识别方法。首先，通过皮尔逊相关性分析从多个随钻参数中筛选出与岩层结构强相关的给进压力、转速、扭矩和泵压4个特征；其次，采用PCA将原始四维特征降维至累计贡献率超95%的3个主成分，消除特征间多重共线性；最后，利用PSO算法对RF分类器超参数进行全局寻优。基于实测数据构建模型并与多种模型对比，结果表明：PCA?PSO?RF融合模型的F1分数高达0.964，精确率与召回率均超过96%，显著优于未经优化的RF模型及主流梯度提升算法；在独立钻孔数据集上验证准确率为84.8%，对无标签数据的预测准确率为83%，其中完整段误判率仅2%，证明了该模型具有良好的鲁棒性与泛化能力。本研究实现了岩层破碎段与完整段的实时高精度识别，为智能钻探与岩土工程勘察提供了可靠的技术支撑。

Abstract:

To address the problems of low efficiency and information lag in conventional core drilling， as well as the limited classification accuracy of existing while-drilling identification models caused by the neglect of parameter multicollinearity and hyperparameter optimization， an intelligent while-drilling identification method for formation structure is proposed based on the fusion of principal component analysis （PCA）， particle swarm optimization （PSO）， and random forest （RF）. First， four features-feed pressure， rotational speed， torque， and pump pressure-that are highly correlated with formation structure are selected from multiple while-drilling parameters using Pearson correlation analysis. Second， PCA is applied to reduce the original four-dimensional features to three principal components with a cumulative contribution rate exceeding 95%， thereby eliminating multicollinearity among the features. Finally， the PSO algorithm is used to globally optimize the hyperparameters of the RF classifier. The model is constructed based on measured data and compared with various other algorithms. The results indicate that the PCA-PSO-RF fusion model achieves an F1-score of 0.964， with both precision and recall exceeding 96%， significantly outperforming the unoptimized RF model and mainstream gradient boosting algorithms. The validation accuracy on an independent borehole dataset is 84.8%， and the prediction accuracy for unlabeled data is 83%， with a misclassification rate of only 2% for intact rock sections， demonstrating the model''s robustness and generalization capability. This study achieves real-time， high-precision identification of fractured and intact rock sections， providing reliable technical support for intelligent drilling and geotechnical engineering investigation.